EP2012260A1 - Procédé et dispositif pour la réalisation d'une mémoire associative basée sur des réseaux neuronaux inhibiteurs - Google Patents
Procédé et dispositif pour la réalisation d'une mémoire associative basée sur des réseaux neuronaux inhibiteurs Download PDFInfo
- Publication number
- EP2012260A1 EP2012260A1 EP07110870A EP07110870A EP2012260A1 EP 2012260 A1 EP2012260 A1 EP 2012260A1 EP 07110870 A EP07110870 A EP 07110870A EP 07110870 A EP07110870 A EP 07110870A EP 2012260 A1 EP2012260 A1 EP 2012260A1
- Authority
- EP
- European Patent Office
- Prior art keywords
- patterns
- matrix
- model
- pattern
- inhibitory
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
- G06N3/082—Learning methods modifying the architecture, e.g. adding, deleting or silencing nodes or connections
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/044—Recurrent networks, e.g. Hopfield networks
Definitions
- the present invention relates to a method and a device for realizing an associative memory using inhibitory neural networks.
- neural networks can compute mappings of the kind f : u ⁇ ⁇ , where u and ⁇ are usually interpreted as activity pattern vectors of the presynaptic and postsynaptic neuron populations, respectively.
- u and ⁇ are usually interpreted as activity pattern vectors of the presynaptic and postsynaptic neuron populations, respectively.
- multilayer neural networks can approximate arbitrary functions f with arbitrary precision (although this may require a very large neuron number in the hidden layers).
- This task is essentially the best match or nearest-neighbor problem and can be extended to implement again continuous mappings f by interpolating between several sample points
- Each address pattern u ⁇ is a binary vector of length m that, on average, contains k one-entries and m - k zero-entries.
- each target pattern v ⁇ is a binary vector of length n that contains l one-entries and n - l zero-entries.
- the patterns are sparse, i.e., k « m and l « n.
- Figure 1 shows the working principle of the binary Willshaw model (for hetero-association).
- M associations between address patterns u ⁇ and content patterns v ⁇ are stored in the binary memory matrix A representing binary synaptic weights of the connection from address population u to content population v .
- all synapses are inactive.
- the synapses are activated according to Hebbian coincidence learning (eq. 1).
- an address pattern ⁇ is propagated through the network.
- v ⁇ (here equal to v 1 )
- a threshold ⁇ is applied.
- yielding a superset v ⁇ D v ⁇ of the original pattern (i.e., no missing one-entries).
- the model may slightly be extended by allowing noisy one-entries and zero-entries in the matrix.
- ⁇ 1 and ⁇ 0 be two binary random m ⁇ n matrices, where each matrix entry is one independently with probability p ⁇ 1 and p ⁇ 0 , respectively.
- the stored information can be retrieved applying an address pattern ⁇ .
- a critical prerequisite for high retrieval quality is the right choice of the threshold value ⁇ : Too low values will lead to high rates of false one-entries (add-errors) whereas too high values will result in high rates of false zero-entries (miss-errors).
- This threshold choice can be simple or rather difficult, depending on the types of errors present in the address.
- c and f denote the number of correct and false one-entries in the address, respectively.
- p 1 the fraction of one-entries in the memory matrix.
- the retrieval or output noise ⁇ may be defined according to the expected Hamming distance d H between v ⁇ and u ⁇ .
- the latter definition is particularly useful when analyzing low quality retrieval for sparse patterns with l ⁇ n .
- the optimal threshold ⁇ opt ⁇ ⁇ 0, 1,... c + f ⁇ can be defined by minimizing ⁇ (or ⁇ T ). For ⁇ ⁇ 1, "high-fidelity" retrieval is obtained.
- C ⁇ : stored information
- #synapses M ⁇ ⁇ T v ⁇ ; v ⁇ ⁇ mn ⁇ M ⁇ ⁇ T l / n , p 01 , p 10 m
- T (v ⁇ ; v ⁇ ⁇ ) is the transinformation (or mutual information) between learned and retrieved content pattern.
- C S synaptic capacity
- information capacity C I making use of the compressibility of the memory matrix for a memory load p 1 ⁇ 0.5
- C S normalizes the stored information to the number of non-silent synapses (instead of all synapses).
- the variable C s is relevant for VLSI implementations of the Willshaw model where each synapse, similar to biology, must be realized in hardware.
- C I normalizes the stored information to the amount of physical memory that is actually required to represent the memory matrix. This is relevant mainly for technical applications where the Willshaw matrix must be maintained in the memory of digital computers .
- equation 17 implies a minimal assembly size k min for the address population. For smaller assemblies it is impossible to achieve high-fidelity ⁇ , even when storing only a single association.
- the two model variants with the naive threshold control (described at the beginning of this section) and the optimal threshold control (eq. 25) will be referred to as the naive and optimal excitatory diluted Willshaw model, respectively.
- a method for forming an associative memory according to the invention uses an inhibitory interpretation of the Willshaw model.
- the excitatory Willshaw model has an equivalent inhibitory interpretation that avoids the problems of the excitatory model described in the previous section.
- memory matrix A eq. 2
- address pattern ⁇ eq. 3
- ⁇ Ap the vector of input activities (eq. 25).
- the excitatory and inhibitory models are actually equivalent.
- the left hand part (2a) of figure 2 shows pattern retrieval in a diluted excitatory network according to the state of the art.
- the right hand part (2b) of figure 2 shows pattern retrieval in a diluted inhibitory network according to a first embodiment of the invention.
- 0-synapses of the excitatory model become -1-synapses, and 1-synapses become 0-synapses.
- the x-axis for the dendritic potentials is shifted such that 0 corresponds to the optimal threshold in each case.
- the plot on the right hand side shows minimal output noise ⁇ (see eq. 10) as a function of the number M of stored associations.
- the retrieval quality of the inhibitory model is generally much better than for the naive excitatory model.
- the plots indicate that, for the same number M of stored pattern associations, the output noise ⁇ (see eq. 10) is much worse for the excitatory model.
- the inhibitory model can store a much larger number of patterns, i.e., it has a much larger storage capacity.
- T was used for the stored (trans-) information which is proportional to M ⁇ .
- C S and C I may be computed (see eqs. 14,15).
- the different capacities for the different model variants are summarized in the table shown below.
- C S performance may be improved by eliminating all silent synapses.
- C S improves by a factor 1/p' 1 ⁇ for the naive excitatory model, and by a factor 1/(1 - p 1 ⁇ ) for the inhibitory model.
- the optimal excitatory model has only factor 1/ p , the same as in eq. 31, because here one cannot eliminate the 0-synapses since they are required to compute the input activity eq. 25.
- the memory matrix interpreted as a bit stream, can be compressed by a Huffman or Golomb code.
- the improvement for the naive excitatory and the inhibitory model is factor 1/ I ( p ' 1 ⁇ ) and 1/ I(p 1 ⁇ ), respectively, but only factor 1/ I (1 - p , 1 -p 1 ⁇ ) for the optimal excitatory model: here a fraction 1 - p of the potential synapses is missing, another fraction 1 - p 1 € are zero, and the remaining synapses are 1.
- the following table shows storage capacity C, synaptic capacity C S , and information capacity C I (see eqs. 13-15) for the different implementations of the Willshaw model (naive excitatory model, optimal excitatory model, and inhibitory model).
- mn is network size
- p ⁇ 1 is connectivity
- p 1 ⁇ is memory load
- T ⁇ denotes the maximal storable information under a given quality constraint ⁇ .
- the asymptotical approximations follow from eq. 31.
- the storable information T' ⁇ and the corresponding memory load p ' 1 ⁇ are usually smaller than for the optimal model variants.
- I (p) and I(p, q) are the information of a binary and ternary random variable, respectively.
- Figure 4 shows asymptotic network capacity C, synaptic capacity C S , and information capacity C 1 as functions of the memory load p 1 ⁇ (see the above table) for different network connectivities p .
- the inhibitory model can achieve large capacities ⁇ 1bps.
- both the excitatory and the inhibitory model can achieve high capacities.
- the excitatory model performs particularly good for sparse potentiation with memory load p 1 ⁇ ⁇ 0 and very sparse patterns k / log n ⁇ 0, while the inhibitory model achieves highest capacities for dense potentiation with p 1 ⁇ ⁇ 1 and moderate sparseness k / log n ⁇ ⁇ .
- the inhibitory model achieves high capacities C S and C I .
- the time complexity of the Willshaw model is essentially the number of synaptic operations necessary to perform a retrieval. Obviously, this is an important performance measure for technical implementations on sequential digital computers.
- a naive retrieval (with uncompressed memory matrix) for an address pattern with z :
- a common way to implement associative memory is simply by a look-up table where each pattern association u ⁇ ⁇ v ⁇ is explicitely stored in one of the table's rows. While the Willshaw model is a two-layer neural network, look-up-tables have an interpretation as three-layer neural networks (see Figure 5 ) with an additional hidden "grandmother cell" layer w.
- each address pattern u ⁇ is stored in a separate row of the M ⁇ m synaptic matrix U describing the connection from u to w.
- each content pattern v ⁇ is stored in a row of the M ⁇ n synaptic matrix V describing the connection from w to v.
- Retrieval is straight forward: An address pattern ⁇ is multiplied with the first synaptic matrix U giving the potentials of layer-w neurons. Then, the winner neuron with the largest potential is selected and the corresponding content pattern is returned as retrieval result.
- the memory matrix is sparse ( p 1 ⁇ ⁇ 0), balanced (0 ⁇ ⁇ ⁇ p 1 ⁇ ⁇ 1 - ⁇ ), or dense ( p 1 ⁇ ⁇ 1) for sub-logarithmic, logarithmic, or supra-logarithmic k ( n ) (see eq. 21).
- the Willshaw model outperforms the cLUT only for dense potentiation (p 1 ⁇ ⁇ 1) if implemented with inhibitory neurons (eq. 36).
- the model proposed in the next section makes use of the inhibitory Willshaw model operating in the most efficient regime of dense potentiation (p 1 ⁇ ⁇ 1) and asymmetric pattern activities ( l « k ).
- Figure 6 shows four layer system for information retrieval according to an embodiment of the invention.
- the first projection from address population u to the first intermediary population w 1 is the inhibitory Willshaw associative memory (IAM).
- the second and third projection from w 1 to another intermediary population w 2 and the content population v is a compressed look-up table (cLUT) as described in the previous section. If the address patterns u ⁇ are not extremely sparse (as is usually the case in applications), then retrieval in this four layer model is much faster than in a compressed look-up table.
- the four-layer architecture introduced above makes use of an inhibitory associative network as the first processing stage. It may be shown that the inhibitory network is more efficient than a compressed look-up table for many parameters, in particular for those required in the four-layer system (i.e., extremely sparse population ⁇ 1 ) . However, depending on the address pattern statistics and the number of pattern associations it may be sometimes better to use a diluted look-up table instead of the inhibitory network. Both model variants have in common that they speed-up retrieval by introducing an intermediary cell layer with very specific and sparse (i.e., grandmother-cell-like) response properties.
- the analysis of the inhibitory Willshaw network relies on a binomial approximation of the retrieval error probability. This approximation can lead to a significant overestimation of the performance of the Willshaw network's performance, in particular for small finite networks or if the address patterns are not sufficiently sparse.
- the intermediary pattern representations of layer w 1 may be optimized in order to minimize error probabilities.
- the patterns w 1 ⁇ are generated at random with size n 1 and activity l 1 (see eq. 40).
- the so-called unit usage of a w 1 neuron is approximately binomially distributed with parameters M and l 1 / n (the unit usage of a neuron is defined as the number of patterns a neuron belongs to).
- the intermediary w 1 patterns arbitrarily, for example, in such a way that the unit usages are approximately equal for all w 1 neurons. This can be realized, for example, by the following greedy-algorithm:
- One possible greedy algorithm for creating a pattern set with approximately uniform synapse usage is the following:
- Figure 7 shows improving storage capacity and retrieval quality of the four-layer system (Fig. ) by using UUU/USU networks.
- the different curves correspond to different methods to construct the patterns in layer w 1 : RND: Standard model with random patterns. Solid line corresponds to the exact theory , crosses (x) to simulations.
- Dashed line corresponds to output noise as predicted by the binomial approximation eq. 16.
- UUU Patterns are constructed using the uniform-unit-usage algorithm UUU (star).
- USU Patterns are constructed using the uniform-synapse-usage algorithm USU (plus).
- USU2/3 Patterns are constructed using two optimized versions of the USU algorithm (triangle, diamond). Note that for these algorithms output noise is much less than predicted by the binomial approximation.
- Figure 7 summarizes data from numerical simulations showing that UUU and USU networks generally perform much better than the standard Willshaw network.
- M the probability of an retrieval error can be reduced by several orders of magnitude.
- M ⁇ increases significantly.
- the UUU/USU algorithms increase pattern capacity by more than factor two, and USU2/3 even more than factor three.
- USU2/3 lead to significantly better results than predicted by the binomial approximation eq.
- the best match problem is also known as nearest-neighbor-search and efficient solutions have widespread applications including the following ones:
Priority Applications (4)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
EP07110870A EP2012260B1 (fr) | 2007-06-22 | 2007-06-22 | Procédé et dispositif pour la réalisation d'une mémoire associative basée sur des réseaux neuronaux inhibiteurs |
JP2010512604A JP5297451B2 (ja) | 2007-06-22 | 2008-06-20 | 抑制性ニューラルネットワークに基づいて連想メモリを実現するための方法および装置 |
PCT/EP2008/005001 WO2009000474A1 (fr) | 2007-06-22 | 2008-06-20 | Procédé et dispositif pour former une mémoire associative sur la base des réseaux neuronaux inhibiteurs |
US12/599,605 US8335752B2 (en) | 2007-06-22 | 2008-06-20 | Method and device for realizing an associative memory based on inhibitory neural networks |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
EP07110870A EP2012260B1 (fr) | 2007-06-22 | 2007-06-22 | Procédé et dispositif pour la réalisation d'une mémoire associative basée sur des réseaux neuronaux inhibiteurs |
Publications (2)
Publication Number | Publication Date |
---|---|
EP2012260A1 true EP2012260A1 (fr) | 2009-01-07 |
EP2012260B1 EP2012260B1 (fr) | 2012-04-11 |
Family
ID=38625856
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
EP07110870A Expired - Fee Related EP2012260B1 (fr) | 2007-06-22 | 2007-06-22 | Procédé et dispositif pour la réalisation d'une mémoire associative basée sur des réseaux neuronaux inhibiteurs |
Country Status (4)
Country | Link |
---|---|
US (1) | US8335752B2 (fr) |
EP (1) | EP2012260B1 (fr) |
JP (1) | JP5297451B2 (fr) |
WO (1) | WO2009000474A1 (fr) |
Cited By (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
EP2259214A1 (fr) * | 2009-06-04 | 2010-12-08 | Honda Research Institute Europe GmbH | Mise en oeuvre d'une mémoire neurale associative basée sur un apprentissage non linéaire de synapses discrètes |
EP2259215A1 (fr) * | 2009-06-04 | 2010-12-08 | Honda Research Institute Europe GmbH | Procédé et structure pour une mémoire neuro-associative basée sur un apprentissage bayésien optimal |
CN103688277A (zh) * | 2011-07-21 | 2014-03-26 | 高通股份有限公司 | 稳健的神经时间编码、学习以及使用振荡来为记忆进行细胞募集的方法和设备 |
Families Citing this family (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US9158847B1 (en) * | 2011-07-19 | 2015-10-13 | Kyndi Inc. | Cognitive memory encoding networks for fast semantic indexing storage and retrieval |
US10268948B2 (en) * | 2015-07-23 | 2019-04-23 | The Boeing Company | Data driven classification and troubleshooting system and method using associative memory and a machine learning algorithm to improve the accuracy and performance of the associative memory |
US11494655B2 (en) | 2017-12-08 | 2022-11-08 | International Business Machines Corporation | Random matrix hardware for machine learning |
US20190286074A1 (en) * | 2018-03-13 | 2019-09-19 | Hrl Laboratories, Llc | Sparse associative memory for identification of objects |
Citations (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US5251269A (en) * | 1987-07-15 | 1993-10-05 | Research Development Corporation Of Japan | Multi-layer neural network modelled after the striate cortex for recognizing visual patterns |
WO2003015026A1 (fr) * | 2001-08-10 | 2003-02-20 | Saffron Technology, Inc. | Neurones artificiels comportant des ponderations qui definissent des projections maximales |
Family Cites Families (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US6366897B1 (en) * | 1999-07-26 | 2002-04-02 | Hnc Software, Inc. | Cortronic neural networks with distributed processing |
US7565491B2 (en) * | 2005-08-04 | 2009-07-21 | Saffron Technology, Inc. | Associative matrix methods, systems and computer program products using bit plane representations of selected segments |
US20100057452A1 (en) * | 2008-08-28 | 2010-03-04 | Microsoft Corporation | Speech interfaces |
-
2007
- 2007-06-22 EP EP07110870A patent/EP2012260B1/fr not_active Expired - Fee Related
-
2008
- 2008-06-20 WO PCT/EP2008/005001 patent/WO2009000474A1/fr active Application Filing
- 2008-06-20 JP JP2010512604A patent/JP5297451B2/ja not_active Expired - Fee Related
- 2008-06-20 US US12/599,605 patent/US8335752B2/en not_active Expired - Fee Related
Patent Citations (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US5251269A (en) * | 1987-07-15 | 1993-10-05 | Research Development Corporation Of Japan | Multi-layer neural network modelled after the striate cortex for recognizing visual patterns |
WO2003015026A1 (fr) * | 2001-08-10 | 2003-02-20 | Saffron Technology, Inc. | Neurones artificiels comportant des ponderations qui definissent des projections maximales |
Non-Patent Citations (5)
Title |
---|
A. KNOBLAUCH, T.WENNEKERS: "Felix++ online documentation , FELIX 2 ANNOTATED LIST", 31 December 2005 (2005-12-31), XP002458565, Retrieved from the Internet <URL:http://web.archive.org/web/20051231112214/http://www.informatik.uni-ulm.de/ni/mitarbeiter/AKnoblauch/kdoc_Felix2/index-long.html> [retrieved on 20071112] * |
ANDREAS KNOBLAUCH: "Synchronization and pattern separation in spiking associative memories and visual cortical areas", October 2003, UNIVERSITÄT ULM, ULM, DE, XP002458566 * |
BUCKINGHAM J ET AL: "On setting unit thresholds in an incompletely connected associative net", NETWORK: COMPUTATION IN NEURAL SYSTEMS, IOP PUBLISHING, BRISTOL, GB, vol. 4, no. 4, 1 November 1993 (1993-11-01), pages 441 - 459, XP020060991, ISSN: 0954-898X * |
KNOBLAUCH ET AL: "Neural associative memory for brain modeling and information retrieval", INFORMATION PROCESSING LETTERS, AMSTERDAM, NL, vol. 95, no. 6, 30 September 2005 (2005-09-30), pages 537 - 544, XP005028095, ISSN: 0020-0190 * |
MOOPENN A ET AL: "ELECTRONIC IMPLEMENTATION OF ASSOCIATIVE MEMORY BASED ON NEURAL NETWORK MODELS", IEEE TRANSACTIONS ON SYSTEMS, MAN AND CYBERNETICS, IEEE INC. NEW YORK, US, vol. SMC-17, no. 2, March 1987 (1987-03-01), pages 325 - 331, XP000743138, ISSN: 0018-9472 * |
Cited By (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
EP2259214A1 (fr) * | 2009-06-04 | 2010-12-08 | Honda Research Institute Europe GmbH | Mise en oeuvre d'une mémoire neurale associative basée sur un apprentissage non linéaire de synapses discrètes |
EP2259215A1 (fr) * | 2009-06-04 | 2010-12-08 | Honda Research Institute Europe GmbH | Procédé et structure pour une mémoire neuro-associative basée sur un apprentissage bayésien optimal |
US8463722B2 (en) | 2009-06-04 | 2013-06-11 | Honda Research Institute Europe Gmbh | Implementing a neural associative memory based on non-linear learning of discrete synapses |
US8521669B2 (en) | 2009-06-04 | 2013-08-27 | Honda Research Institute Europe Gmbh | Neural associative memories based on optimal bayesian learning |
CN103688277A (zh) * | 2011-07-21 | 2014-03-26 | 高通股份有限公司 | 稳健的神经时间编码、学习以及使用振荡来为记忆进行细胞募集的方法和设备 |
CN103688277B (zh) * | 2011-07-21 | 2016-08-17 | 高通股份有限公司 | 稳健的神经时间编码、学习以及使用振荡来为记忆进行细胞募集的方法和设备 |
Also Published As
Publication number | Publication date |
---|---|
EP2012260B1 (fr) | 2012-04-11 |
US8335752B2 (en) | 2012-12-18 |
JP5297451B2 (ja) | 2013-09-25 |
WO2009000474A1 (fr) | 2008-12-31 |
JP2010531030A (ja) | 2010-09-16 |
US20100306145A1 (en) | 2010-12-02 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
EP2012260B1 (fr) | Procédé et dispositif pour la réalisation d'une mémoire associative basée sur des réseaux neuronaux inhibiteurs | |
EP2259214B1 (fr) | Mise en oeuvre d'une mémoire neurale associative basée sur un apprentissage non linéaire de synapses discrètes | |
EP2259215B1 (fr) | Procédé et structure pour une mémoire neuro-associative basée sur un apprentissage bayésien optimal | |
Ou et al. | Multi-class pattern classification using neural networks | |
CA3039386A1 (fr) | Modele de reseau neuronal profond pour le traitement de donnees par l'intermediaire de hierarchies de taches linguistiques multiples | |
Zhou et al. | Discriminative deep belief networks for image classification | |
US20190286980A1 (en) | Cluster compression for compressing weights in neural networks | |
Furber et al. | Sparse distributed memory using N-of-M codes | |
EP4035273B1 (fr) | Conception et formation de neurones binaires et de réseaux neuronaux binaires avec des codes de correction d'erreur | |
Castro et al. | Data-partitioning using the hilbert space filling curves: Effect on the speed of convergence of Fuzzy ARTMAP for large database problems | |
Slakey et al. | Encoding categorical variables with conjugate bayesian models for wework lead scoring engine | |
Vivekanandan | A fast genetic algorithm for mining classification rules in large datasets | |
CN111581386A (zh) | 多输出文本分类模型的构建方法、装置、设备及介质 | |
Sacramento et al. | Tree-like hierarchical associative memory structures | |
Yellasiri et al. | DECISION TREE INDUCTION USING ROUGH SET THEORY--COMPARATIVE STUDY. | |
Knoblauch | Neural associative memory and the Willshaw–Palm probability distribution | |
He et al. | Semisupervised network embedding with differentiable deep quantization | |
Shetty et al. | Comparative analysis of different classification techniques | |
Dubitzky et al. | Data mining and machine learning methods for microarray analysis | |
Mariette et al. | Accelerating stochastic kernel SOM | |
Hsu et al. | On minimizing diagonal block-wise differences for neural network compression | |
Wang et al. | Optimal cluster selection based on Fisher class separability measure | |
CN115391414B (zh) | 一种基于大数据的银行市场拓展系统及方法 | |
Wolcott et al. | Scalable record linkage | |
Aldana-Bobadilla et al. | A novel data reduction method based on information theory and the Eclectic Genetic Algorithm |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PUAI | Public reference made under article 153(3) epc to a published international application that has entered the european phase |
Free format text: ORIGINAL CODE: 0009012 |
|
AK | Designated contracting states |
Kind code of ref document: A1 Designated state(s): AT BE BG CH CY CZ DE DK EE ES FI FR GB GR HU IE IS IT LI LT LU LV MC MT NL PL PT RO SE SI SK TR |
|
AX | Request for extension of the european patent |
Extension state: AL BA HR MK RS |
|
17P | Request for examination filed |
Effective date: 20090122 |
|
17Q | First examination report despatched |
Effective date: 20090311 |
|
AKX | Designation fees paid |
Designated state(s): DE FR GB |
|
GRAJ | Information related to disapproval of communication of intention to grant by the applicant or resumption of examination proceedings by the epo deleted |
Free format text: ORIGINAL CODE: EPIDOSDIGR1 |
|
GRAP | Despatch of communication of intention to grant a patent |
Free format text: ORIGINAL CODE: EPIDOSNIGR1 |
|
GRAS | Grant fee paid |
Free format text: ORIGINAL CODE: EPIDOSNIGR3 |
|
GRAA | (expected) grant |
Free format text: ORIGINAL CODE: 0009210 |
|
AK | Designated contracting states |
Kind code of ref document: B1 Designated state(s): DE FR GB |
|
REG | Reference to a national code |
Ref country code: GB Ref legal event code: FG4D |
|
REG | Reference to a national code |
Ref country code: DE Ref legal event code: R096 Ref document number: 602007021845 Country of ref document: DE Effective date: 20120606 |
|
PLBE | No opposition filed within time limit |
Free format text: ORIGINAL CODE: 0009261 |
|
STAA | Information on the status of an ep patent application or granted ep patent |
Free format text: STATUS: NO OPPOSITION FILED WITHIN TIME LIMIT |
|
26N | No opposition filed |
Effective date: 20130114 |
|
REG | Reference to a national code |
Ref country code: DE Ref legal event code: R097 Ref document number: 602007021845 Country of ref document: DE Effective date: 20130114 |
|
REG | Reference to a national code |
Ref country code: DE Ref legal event code: R084 Ref document number: 602007021845 Country of ref document: DE Effective date: 20140711 |
|
REG | Reference to a national code |
Ref country code: GB Ref legal event code: 746 Effective date: 20150330 |
|
REG | Reference to a national code |
Ref country code: FR Ref legal event code: PLFP Year of fee payment: 9 |
|
PGFP | Annual fee paid to national office [announced via postgrant information from national office to epo] |
Ref country code: GB Payment date: 20150623 Year of fee payment: 9 |
|
PGFP | Annual fee paid to national office [announced via postgrant information from national office to epo] |
Ref country code: FR Payment date: 20150623 Year of fee payment: 9 |
|
PGFP | Annual fee paid to national office [announced via postgrant information from national office to epo] |
Ref country code: DE Payment date: 20150826 Year of fee payment: 9 |
|
REG | Reference to a national code |
Ref country code: DE Ref legal event code: R119 Ref document number: 602007021845 Country of ref document: DE |
|
GBPC | Gb: european patent ceased through non-payment of renewal fee |
Effective date: 20160622 |
|
REG | Reference to a national code |
Ref country code: FR Ref legal event code: ST Effective date: 20170228 |
|
PG25 | Lapsed in a contracting state [announced via postgrant information from national office to epo] |
Ref country code: FR Free format text: LAPSE BECAUSE OF NON-PAYMENT OF DUE FEES Effective date: 20160630 Ref country code: DE Free format text: LAPSE BECAUSE OF NON-PAYMENT OF DUE FEES Effective date: 20170103 |
|
PG25 | Lapsed in a contracting state [announced via postgrant information from national office to epo] |
Ref country code: GB Free format text: LAPSE BECAUSE OF NON-PAYMENT OF DUE FEES Effective date: 20160622 |